I created a small SIFT application that captures key points and saves it in a text file. I use this to capture information from a logo (e.g. AT & T) and use it to compare with other images with this logo. The problem is that many of my images have logo options that, due to scaling, rotation, or lighting, do not raise it. I was wondering if it is possible to get a set of images, capture its key points and run it through some kind of training algorithm to improve detection.
I searched the Internet for ways to teach SIFT key points, but they are all in some kind of phd paper, which is included in all these mathematical algorithms, which, to be honest, they throw me away, because I did not take the math class for a while.
If anyone has any tips or links to understand how the training works or what needs to be done to implement it, let me know. Or, if someone has easier ways to do this without SIFT, I would really appreciate other forms of detection. Below is a list of what I have tried:
- Surf
- Failure to return incorrect results
- Haar Characteristics with Adaboosting
- Failed since I started training 100 positive models with 100 negative images on 11/11/2011 and still working as of 7/19/2011.
- Template Correspondence to various transformations of the same logo with and without a threshold
- It failed, as I would have to exponentially create logos based on the number of times when he could not detect any images.
Thanks in advance
source share